Guest

Cisco Service Adapters

Performance Study: CSA on Cisco 7200 and 7500 Series Routers

 

CSA Based Data Sheet

Cisco Compression Service Adapter Performance Study for Cisco 7200 and 7500 Series Routers


This study compares the performance of software-based (Cisco IOS®) data compression as well as Compression Service Adapter (CSA) based hardware compression on both Cisco 7200 and 7500 series router platforms. The compression algorithm used in this study is LZS—STAC's lossless data compression algorithm.

Cisco IOS software supports three different methods to implement compression.

  • Software-based compression using the main route processor (Route Switch Processor) [RSP] or Network Processing Engine [NPE])

  • Software-based compression using Versatile Interface Processors (VIPs) (Distributed Processor based)

  • Hardware-based compression using Compression Service Adapter (CSA)

Executive Summary

This performance study concludes that software-based compression is a cost-effective solution for low to moderately high link-speed connections with larger data packet sizes. Furthermore, software-based compression can significantly improve overall link efficiency and available bandwidth for single T1 or E1 links, depending on average packet size and processor/VIP selections. Other CPU activities must be accounted for such as security, quality of service, and so on, in estimating the true compression performance in software.

Compression Service Adapter (CSA) based hardware compression is the solution of choice for high link speed connections with both small and large packet sizes. The CSA performed well on all link speeds under test, including 8 Mbps full duplex. Although not part of this test, the CSA may also be used with the High-Speed Serial Interface (HSSI) port adapter with link speeds approaching 16 Mbps full duplex before compression efficiency is impacted.

Hardware-based compression also benefits from the fact that it does not depend on the processor power as much as software compression does. CSA based compression off-loads the main processor of all tasks related to compression and decompression so it can concentrate only on efficiently executing the main router functions.

Performance Study

This study is done in three parts. First, the throughput of various links is measured without any compression enabled. Second, the software-based (Cisco IOS) compression is enabled. This test, besides giving the throughput performance, also provides the percentage CPU load and the compression ratio for the test data. The last test performed is hardware-based (CSA) compression, which also provides the throughput performance, percentage CPU load, and the compression ratio.

The tests are conducted for 103 byte and 55 byte size packets. The 103 byte packet represents a typical small-sized packet, whereas a 551 byte packet represents a default Trivial File Transfer Protocol (TFTP) packet. The data payload was developed out of the Calgary Corpus, a standard payload used by the UNIX community for testing compression algorithms. The compression algorithm used in all the cases is STAC's-LZS lossless data compression algorithm is used in all cases.

Both hardware -assisted and software-based compression performances are measured for different link speeds: 64 kbps and T1, including 1.54, 2.015, 4.030, and 8.061Mbps. For this test, a PA-4T+ (1.54 Mbps) card was used for one PPP link. The remaining three ports were administratively shut off.

Results are computed for the Cisco 7200 and 7500 platforms. In the Cisco 7200 platform, the compression performance is measured against both models of NPEs (NPE-150 and NPE-200). In the Cisco 7500 platform, the compression performance is measured against RSP 4 and both VIPs (VIP2-40 and VIP2-50).

The following performance charts show the maximum number of data packets (measured in thousands of packets per second) for both 551 byte- and 103 byte-sized packets, accommodated for a given link speed (typically measured in megabits per second).

The uncompressed packet rate reflects the number of packets per second required to achieve line rate for various link speeds. Compression technology reduces packet size, thereby increases the number of packets transmitted for the same link rate. The effect is virtually increasing overall link bandwidth.

For example, in test 1, for a 4Mbps link speed, the maximum number of uncompressed packets (551 byte) that can be transported is 0.914kpps. This number translates to 4 Mbps, which is also the maximum link rate. Using software compression, the maximum number of packets (551 byte) transported is 1.156 Kpps. This number translates to 5.095 Mbps of data transported on the same physical link operating at the same link speed, which is 4 Mbps. This represents an increase of approximately 1Mbps of data throughput. Now, using CSA based hardware compression, the maximum number of packets (551 byte) transported is 2.062kpps. This number translates to 9.089Mbps of data transported on the same 4Mbps link, more than twice compared to uncompressed data transport.

Case 1

Platforms: VIP2-40 in Cisco 7500 platform

Port Adapter: PA-4T+ and CSA-Comp/1 (768 KB)

Cisco IOS Ver.: 11.1(14.5) CA

Data Packet Size: 551 bytes and 103 bytes

Figure 1 shows that compression definitely helps to increase the data throughput, with hardware compression sending more data on the link than its software counterpart. Compressing data packets and thereby better utilizing the available WAN link bandwidth translates directly into WAN access cost savings.

In the case of software compression, the CPU performance is the bottleneck for compressing high-speed links (see 8M case). The fact that the CPU utilization was near 99 percent, which is why the software-compressed throughput is below the uncompressed throughput. The compression ratio is 2.35.

The link utilization with hardware compression is more than twice that for link speeds up to 8M, which is equivalent to four E1 links. The CPU utilization of the VIP Processor supports the CSA port adapter is be less than 50. The compression ratio is 2.28.

The software compression performance depends heavily on the processor power. The smaller packet size puts more load on the processor, as evident from the reduced compression performance for higher-speed links, 4M for instance.

CSA (hardware-assisted compression) is definitely recommended for compressing high-speed links. This compression also relieves the processor for other important router functions such as generating/updating routing tables and so on.


Figure 1: Compression Performance in VIP2-40 (551 byte packet)


Link Speeds 64K T1 2M 4M 8M
Uncompressed

0.064

1.607

2.011

4.021

8.052

Software Compressed

0.128

3.713

4.643

5.097

7.215

CSA Compressed

0.128

3.603

4.507

9.093

18.070

Maximum data throughput in Mbps for various link speeds



Figure 2: Compression Performance in VIP2


Link Speeds 64k T1 2M 4M 8M
Uncompressed

0.063

1.604

2.002

4.011

8.025

Software Compressed

0.106

2.801

2.903

2.914

2.923

CSA Compressed

0.106

2.759

3.450

6.900

13.699

Maximum data throughput in Mbps for various link speeds


Case 2

Platforms: VIP2-50 in Cisco 7500 platform

Port Adapter: PA-4T+ and CSA-Comp/1 (768 KB)

Cisco IOS Ver.: 11.1(14.5) CA

Data Packet Size: 551 bytes and 103 bytes

Software compression performance is better with the VIP2-50 than with the VIP2-40 because the VIP2-50 has a faster processor. This is evident for higher-speed links such as 4 Mbps (two E1 links) seen in Figure 3. The processor load was near 99 percent for link speeds of 4 Mbps and beyond. This once again confirms that software compression is ideal for low-speed links. The compression ratio in this case is 2.35.

The link utilization with hardware compression is more than twice as much as that seen in Figure 1. Effective use of the WAN links with hardware compression results in an order of magnitude higher link utilization and thereby reduces the WAN link access costs. The CPU utilization of the VIP that supports the CSA port adapter was less than 50 percent. The compression ratio is 2.28.

The software compression performance depends heavily on the processor power. The smaller packet size puts more load on the processor, as evident from the reduced compression performance for higher speed links, like 4 Mbps and 8 Mbps for instance. The processor utilization was near 99 percent in both those cases. The compression ratio is 1.81.


Figure 3: Compression Performance in VIP2-50 (551 byte packet)


Link Speeds 64k T1 2M 4M 8M
Uncompressed

0.064

1.607

2.011

4.026

8.061

Software Compressed

0.128

3.713

4.643

7.88

7.713

CSA Compressed

0.128

3.603

4.507

9.093

18.07

Maximum data throughput in Mbps for various link speeds



Figure 4: Compression Performance in VIP2-50 (103 byte packet)


Link Speeds 64k T1 2M 4M 8M
Uncompressed

0.063

1.604

2.014

4.029

8.025

Software Compressed

0.106

2.801

3.511

4.696

4.861

CSA Compressed

0.106

2.759

3.45

6.9

13.802



Case 3

Platforms: RSP4 in Cisco 7500 platform

Cisco IOS Ver.: 11.1(14.5) CA

Data Packet Size: 551 bytes and 103 bytes

Figure 5 shows the impact of compressing a stream of small packets as compared with packets with software compression (RSP4 based). Large-sized packets compress better because they have more data that can be compressed. Smaller sized packets did not exhibit this advantage.

Moreover, software based compression addresses processor bottlenecks. As seen in Figure 5, there is performance degradation in using RSP4 based software compression for small-sized data packets in high-speed links results in performance degradation. Larger sized data packets fared well here, although with reduced performance because of higher processor utilization (99 percent in this case).

Another reason that RSP-based compression performance degrades for high-speed links with small-sized packets is that they contain less compressible data compared to large sized packets. Packet processing frequency also increases with smaller packet sizes. In addition, both of these happen at the process level, regardless of what switching mode is configured. Again, in software compression, packets will always take the process switching path, and hence the high CPU utilization.


Figure 5: Software Compression in RSP4 (551 & 103 byte packets


Link Speeds 64K T1 2M 4M 8M
Uncompressed (103 byte packet)

0.063

1.604

2.002

4.011

8.025

Compressed (103 byte packet)

0.114

2.801

3.511

4.263

4.264

Uncompressed (551 byte packet)

0.064

1.607

2.011

4.021

8.052

Compressed (551 byte packet)

0.157

3.757

4.696

9.371

11.345

Maximum data throughput in Mbps for various link speeds


Case 4

Platforms: NPE-150 in Cisco 7200 platform

Port Adapter: PA-4T+ and CSA-Comp/1 (768 kB)

Cisco IOS Ver.: 11.1(14.5) CA

Data Packet Size: 551 bytes and 103 bytes

The software compression performance was nearly flat for high-speed links beyond T1 because the processor utilization was at 99 percent. The compression ratio is 2.35. The processor in the NPE-150 executes the normal routing functions, such as routing table generation/update, in addition to compression/decompression.

The link utilization with hardware compression is more than twice that for link speeds up to 8M, which is equivalent to four E1 links. The CPU utilization was less than 50 percent. The compression ratio is 2.28. This points to the benefits of using hardware-assisted (CSA) compression for all link speeds in the Cisco 7200 platform.

The software compression performance was nearly flat for high-speed links beyond T1 because the processor utilization is over 90 percent. The compression ratio is 1.78. The processor in the NPE-150 executes the normal routing functions, such as routing table generation/update in addition to compression/decompression. This shows that software compression is not an ideal choice if the data stream is made up of small-sized packets and the link is a high-speed link.

The link utilization with hardware compression is definitely better than that with the software compression. The CPU utilization increases with higher link speeds. The compression ratio is 2.28. This points to the benefits of using hardware-assisted (CSA) compression for all link speeds in the Cisco 7200 platform.


Figure 6: Compression Performance in NPE-150 (551 byte packet)


Link Speeds 64k T1 2M 4M 8M
Uncompressed

0.064

1.612

2.015

4.03

8.052

Software Compressed

0.153

3.744

4.41

4.392

4.407

CSA Compressed

0.128

3.603

4.52

9.022

17.969

Maximum data throughput in Mbps for various link speeds



Figure 7: Compression Performance in NPE-150 (103 byte packet)


Link Speeds 64k T1 2M 4M 8M
Uncompressed

0.064

1.606

2.008

4.016

8.034

Software Compressed

0.112

1.202

1.446

1.454

1.454

CSA Compressed

0.106

2.759

3.45

6.9

12.051

Maximum data throughput in Mbps for various link speeds


Case 5

Platforms: NPE-200 in Cisco 7200 platform

Port adapter: PA-4T+ and CSA-Comp/1 (768 KB)

Cisco IOS Ver.: 11.1(14.5) CA

Data packet size: 551 bytes and 103 bytes

Software compression performance in the NPE-200 is better than in the NPE-150 (as shown in Figure 8) because the NPE-200 has a faster processor. The processor in the NPE-200 executes the normal routing functions, such as routing table generation/update, in addition to compression/decompression. The software compression performance was nearly flat for high-speed links beyond 2M because the processor utilization was at 99 percent. The compression ratio is 2.36.

The link utilization with hardware compression is more than twice that for link speeds up to 8M, which is equivalent to four E1 links. The CPU utilization was less than 30 percent. The compression ratio is 2.28. The results point to the benefits of using hardware-assisted (CSA) compression for all link speeds in the Cisco 7200 platform.

The software compression performance of the NPE-200, although nearly flat for high-speed links beyond T1, is better than that of the NPE-150 results because the NPE-200 has a higher-speed processor. The software compressed data throughput did not improve for links beyond T1 because the processor utilization is over 90 percent. The compression ratio is 1.78. The processor in the NPE-200 executes the normal routing functions, such as routing table generation/update in addition to compression and decompression. This results that software compression is not an ideal choice if the data stream is made up of small-sized packets and the link is a high-speed link.

The link utilization with hardware compression is definitely better than that with the software compression. The CPU utilization increases with higher link speeds. The compression ratio is 2.28. This points to the benefits of using hardware-assisted (CSA) compression for all link speeds in the Cisco 7200 platform.


Figure 8: Compression Performance in NPE-200 (551 byte packet)


Link Speeds 64k T1 2M 4M 8M
Uncompressed

0.064

1.612

2.015

4.03

8.052

Software Compressed

0.153

3.744

4.683

6.063

6.157

CSA Compressed

0.128

3.603

4.52

9.022

17.969

Maximum data throughput in Mbps for various link speeds



Figure 9: Compression Performance in NPE-200 (103 byte packet)


Link Speeds 64k T1 2M 4M 8M
Uncompressed

0.064

1.606

2.008

4.016

8.034

Software Compressed

0.112

2.372

2.317

2.245

2.226

CSA Compressed

0.106

2.759

3.45

6.9

12.051

Maximum data throughput in Mbps for various link speeds